Re: [PATCH v2 00/15] x86/boot: Rework PE header generation

[Dave Young]

On Tue, 24 Oct 2023 at 01:37, Jan Hendrik Farr <kernel@xxxxxxxx> wrote:
>
> On 23 13:22:53, Ard Biesheuvel wrote:
> > On Tue, 3 Oct 2023 at 04:03, Jan Hendrik Farr <kernel@xxxxxxxx> wrote:
> > >
> > > On 12 09:00:51, Ard Biesheuvel wrote:
> > > > From: Ard Biesheuvel <ardb@xxxxxxxxxx>
> > > >
> > > > Now that the EFI stub boot flow no longer relies on memory that is
> > > > executable and writable at the same time, we can reorganize the PE/COFF
> > > > view of the kernel image and expose the decompressor binary's code and
> > > > r/o data as a .text section and data/bss as a .data section, using 4k
> > > > alignment and limited permissions.
> > > >
> > > > Doing so is necessary for compatibility with hardening measures that are
> > > > being rolled out on x86 PCs built to run Windows (i.e., the majority of
> > > > them). The EFI boot environment that the Linux EFI stub executes in is
> > > > especially sensitive to safety issues, given that a vulnerability in the
> > > > loader of one OS can be abused to attack another.
> > >
> > > This split is also useful for the work of kexecing the next kernel as an
> > > EFI application. With the current EFI stub I have to set the memory both
> > > writable and executable which results in W^X warnings with a default
> > > config.
> > >
> > > What made this more confusing was that the flags of the .text section in
> > > current EFI stub bzImages are set to
> > > IMAGE_SCN_MEM_EXECUTE | IMAGE_SCN_MEM_READ. So if you load that section
> > > according to those flags the EFI stub will quickly run into issues.
> > >
> > > I assume current firmware on x86 machines does not set any restricted
> > > permissions on the memory. Can someone enlighten me on their behavior?
> > >
> >
> > No current x86 firmware does not use restricted permissions at all.
> > All memory is mapped with both writable and executable permissions,
> > except maybe the stack.
> >
> > The x86 Linux kernel has been depending on this behavior too, up until
> > recently (fixes are in -rc now for the v6.6 release). Before this, it
> > would copy its own executable image around in memory.
> >
> > So EFI based kexec will need to support this behavior if it targets
> > older x86 kernels, although I am skeptical that this is a useful
> > design goal.
>
> I don't see this as an important goal either.
>
> > I have been experimenting with running the EFI stub code in user space
> > all the way until ExitBootServices(). The same might work for UKI if
> > it is layered cleanly on top of the EFI APIs (rather than poking into
> > system registers or page tables under the hood).
> >
> > How this would work with signed images etc is TBD but I quite like the
> > idea of running everything in user space and having a minimal
> > purgatory (or none at all) if we can simply populate the entire
> > address space while running unprivileged, and just branch to it in the
> > kexec() syscall. I imagine this being something like a userspace
> > helper that is signed/trusted itself, and gets invoked by the kernel
> > to run EFI images that are trusted and tagged as being executable
> > unprivileged.
>
> I've been experimenting with running EFI apps inside kernel space instead.
> This is the more natural approach for signed images. Sure, a malicious EFI
> app could do arbitrary stuff in kernel mode, but they're signed. On the other
> hand running this entirely in user space would at least guarantee that the
> system can not crash due to a misbehaving EFI app (at least until
> ExitBootServices()).
>
> The question of whether or not to make this the job of a userspace helper that
> is signed must have come up when kexec_file_load syscall was added. It would
> have also been an option at the time to extend trust to a signed version of
> the userspace kexec tool.
>
> Why was kexec_file_load created instead of restricting kexec_load to a signed
> version of the kexec userspace tool?

I think one of the reasons is that it is hard to handle dynamic linked
libraries, not only the kexec-tools binary.

Thanks
Dave